MiniMax M2.7
MiniMax-M2.7 is a next-generation large language model built for autonomous real-world productivity and continuous improvement. Through advanced agentic capabilities and multi-agent collaboration, it can plan, execute, evaluate, and refine complex tasks in dynamic environments while actively contributing to its own evolution. Optimized for production-grade workflows, M2.7 excels at live debugging, root-cause analysis, financial modeling, and end-to-end document creation across Word, Excel, and PowerPoint. It achieves 56.2% on SWE-Pro, 57.0% on Terminal Bench 2, and a 1495 ELO rating on GDPval-AA, establishing a new benchmark for multi-agent systems operating in real-world digital workflows.
Related Models
MiniMax M3
MiniMax-M3 is MiniMax’s latest M-series multimodal foundation model, built for agentic reasoning, tool use, coding, and long-context tasks. It supports text, image, and video inputs with text output, offering a 1M-token context window, extended thinking, function calling, and structured outputs. With strong capabilities in long-horizon agent workflows, software development, multimodal understanding, and extended response generation, MiniMax-M3 is ideal for autonomous agents, coding assistants, document and video analysis, and production-grade applications that require massive context at a competitive cost.
Read Me
MiniMax M2.7 API
MiniMax-M2.7 is a text reasoning model built for real-world software engineering, elaborate agent harnesses, and professional productivity work. MiniMax describes it as the first model to participate deeply in its own development cycle, where an internal version helped analyze failures, modify agent scaffolds, run evaluations, and improve reinforcement-learning workflows.
On iCreat, you can call MiniMax-M2.7 through an OpenAI-compatible Chat Completions API. The endpoint supports standard system and user messages, streamed responses, and thinking configuration, giving applications that already use the OpenAI SDK a direct path to test the model.
Model Positioning
MiniMax-M2.7 extends the M2 series from efficient coding and tool use into more complex, system-level work. Its focus includes complete project delivery, production incident analysis, machine-learning workflows, multi-round office editing, and Agent Teams coordinated through an application-provided harness.
It is a practical option for teams that want an open-weight reasoning model with strong agentic specialization and a relatively small active-parameter footprint. It is not a multimodal model, and applications that need native image understanding should select a model with official visual input support.
Key Capabilities
Real-World Software Engineering
MiniMax-M2.7 targets more than isolated code generation. It can support log analysis, bug diagnosis, refactoring, code-security review, repository-level reasoning, and workflows that connect monitoring evidence with code and deployment history.
Complex Agent Harnesses
The model was developed around long-horizon agent trajectories and can work with complex Skills, dynamic tool discovery, memory, and multi-agent role structures. The application still supplies the tools, permissions, state, and orchestration layer.
Machine-Learning Workflows
M2.7 was used inside MiniMax to help review experiment results, inspect logs, propose scaffold changes, launch evaluations, and decide whether to keep or revert modifications. This makes it relevant to research engineering and iterative ML operations when paired with controlled execution tools.
Professional Office Delivery
MiniMax emphasizes high-fidelity, multi-round work across Word, Excel, and PowerPoint tasks. With the appropriate document tools, the model can help revise structured business artifacts while following complex Skills and feedback across multiple turns.
Long-Context Text Reasoning
The official MiniMax API exposes a 204,800-token context window for M2.7. This supports large code and document inputs, while the standard model is listed at approximately 60 output tokens per second.
Pricing
| Token Type | Price |
|---|---|
| Input | $0.30 per 1M tokens |
| Output | $1.20 per 1M tokens |
Use Cases
- Production incident investigation: correlate logs, deployment changes, traces, and repository evidence to develop testable root-cause hypotheses.
- Repository-scale engineering: plan and execute multi-file fixes, refactors, reviews, and full-project delivery through a controlled coding harness.
- ML research operations: assist with experiment setup, log inspection, metric analysis, evaluation loops, and scaffold changes.
- Office artifact workflows: revise reports, spreadsheets, and presentations across multiple feedback rounds using application-provided file tools.
- Agent and Skill systems: serve as the reasoning model behind coordinated agents, persistent memory, complex Skills, and dynamic tool selection.
Every workflow that changes files, runs code, queries systems, or edits business artifacts still requires explicit tools, scoped permissions, validation, and approval rules.
Model Comparison
MiniMax-M2.7 vs MiniMax-M2.5
| Factor | MiniMax-M2.7 | MiniMax-M2.5 |
|---|---|---|
| Context Window | 204,800 tokens | 204,800 tokens |
| Standard Output Speed | Approximately 60 tokens/s | Approximately 60 tokens/s |
| Positioning | Complex agent harnesses and system-level execution | Coding, tool use, search, and real-world productivity |
| Engineering Scope | Incidents, ML workflows, code security, full projects | Full-lifecycle coding and application delivery |
| Agent Direction | Agent Teams, complex Skills, dynamic tool search | Agentic tool use and search |
MiniMax-M2.7 vs Claude Sonnet 4.6 and GPT-5.4
| Factor | MiniMax-M2.7 | Claude Sonnet 4.6 | GPT-5.4 |
|---|---|---|---|
| Positioning | Open-weight agentic engineering and professional work | Balanced coding, computer use, and knowledge work | Frontier coding and complex professional work |
| Context Window | 204,800 tokens | 1M tokens | 1,050,000 tokens |
| Maximum Output | Not separately stated | 64K tokens | 128K tokens |
| Official Input Modalities | Text | Text and image | Text and image |
| Reasoning Control | Thinking enabled; iCreat example uses low reasoning effort |
Adaptive Thinking; low, medium, high, max effort |
none, low, medium, high, xhigh reasoning effort |
| Best Fit | Custom agent harnesses, engineering, ML and office workflows | Claude-based long-context agents and visual document work | OpenAI tool ecosystem and very long professional workflows |
Why Choose MiniMax-M2.7?
Choose MiniMax-M2.7 when the workload is primarily text-based and needs a model specialized for real engineering and agent execution. Its combination of long-context reasoning, efficient MoE activation, tool-oriented training, and strong professional-work focus makes it suitable for custom agent platforms rather than only conversational use.
Through iCreat, teams can evaluate that capability with an OpenAI-compatible endpoint without operating the model's large self-hosting footprint. The Playground is useful for prompt-level testing, while the API is the appropriate path for integrating tools, state, and verification into a production workflow.
Specifications
| Category | Description |
|---|---|
| Model Name | MiniMax-M2.7 |
| Provider | MiniMax |
| Model ID | MiniMax-M2.7 |
| Release Date | March 18, 2026 |
| Model Type | Text reasoning Mixture-of-Experts LLM |
| Context Window | 204,800 tokens |
| Maximum Output | Not separately stated by MiniMax |
| Official Input Modalities | Text |
| Output Modalities | Text |
| Knowledge Cutoff | Not published by MiniMax |
| M2-Series Parameters | 229.9B total; approximately 9.8B activated per token |
| Official Standard Output Speed | Approximately 60 tokens/s |
| iCreat Reasoning Configuration | Thinking enabled; documented example uses low reasoning effort |
| iCreat API Features | OpenAI-compatible Chat Completions API and streaming |
| Best Suited For | Software engineering, agent harnesses, ML workflows, office productivity |
Architecture
The MiniMax M2 series uses a 62-layer decoder-only Mixture-of-Experts Transformer with 229.9B total parameters and approximately 9.8B activated for each token. The published design includes 256 fine-grained experts, grouped-query attention, and a Multi-Token Prediction module that also supports speculative decoding.
M2.7's main differentiation comes from agent-native post-training rather than a newly claimed parameter scale. MiniMax trained the series with executable workspaces, verifiable trajectories, artifact-aligned rewards, and the Forge reinforcement-learning system, then used M2.7 in iterative scaffold-improvement experiments.
Notes
Use streaming for long reasoning and agent tasks so applications can surface progress instead of waiting for a complete response. Track completion and reasoning usage together when setting request budgets.
Treat the 204,800-token context window as a shared request envelope. Leave room for system instructions, tool definitions, conversation state, reasoning, and the final response rather than filling the entire window with source material.
For tool-driven work, persist the task goal, completed actions, verified results, and unresolved decisions outside the model. Use least-privilege credentials and require approval before destructive, financial, or production-changing actions.
Evaluate end-to-end completion rather than benchmark labels alone. Measure tool-call validity, recovery from failed actions, artifact quality, latency, token usage, and the amount of human correction required.
FAQ
Does “self-evolution” mean MiniMax-M2.7 updates itself during an iCreat API request?
No. MiniMax uses the term for development experiments in which an internal M2.7 agent helped analyze training failures, modify scaffolds, run evaluations, and improve workflows. A normal API request does not retrain the model or change its weights.
Does MiniMax-M2.7 automatically launch Agent Teams and tools?
No. Agent Teams, Skills, memory, and dynamic tool search require an application-provided agent harness. The model supplies reasoning and tool-selection capability, while the application controls execution and permissions.
Is MiniMax-M2.7 the same as MiniMax-M2.7-highspeed?
MiniMax describes them as having the same model performance but different serving speed. The iCreat model on this page is MiniMax-M2.7, so do not assume the highspeed serving profile applies to this endpoint.
What is MiniMax-M2.7's knowledge cutoff?
MiniMax has not published one. For current facts, connect the application to a trusted retrieval or search source and require citations or evidence appropriate to the task.